This Week in Motion AI: Muscles Enter the Loop

Week of 9–15 June 2026 — the field reaches below the skeleton, training AI on the muscles that actually produce movement

1. Muscle-Actuated Humanoids at Scale: MuscleMimic

Towards Embodied AI with MuscleMimic: Unlocking Full-Body Musculoskeletal Motor Learning at Scale Li, C., Wang, C., Ziliotto, B., Simos, M., Kovecses, J., Durandau, G., & Mathis, A. (EPFL / McGill). arXiv:2603.25544. https://arxiv.org/abs/2603.25544

The problem: Almost all AI motion generation operates at the level of the skeleton — joint positions and rotations over time. But the skeleton is not what produces movement. Muscles produce movement, and the relationship between muscle activation and resulting motion is complex, nonlinear, and redundant (many muscle activation patterns can produce the same joint trajectory). Training AI on skeletal motion skips the layer where effort, intention, and quality actually live. The obstacle to working at the muscle level has been computational: biomechanically accurate muscle simulation is extraordinarily expensive, and validated open full-body muscle models have been scarce.

The approach: MuscleMimic provides two validated musculoskeletal models — a 126-muscle upper-body model for bimanual manipulation and a 416-muscle full-body model for locomotion — together with a pipeline that retargets standard SMPL-format motion capture onto these muscle-actuated structures while preserving kinematic and dynamic consistency. Using massively parallel GPU simulation, the framework achieves order-of-magnitude speedups over prior CPU-based muscle simulation, enabling a single generalist policy to learn hundreds of diverse motions within days, and to fine-tune to novel motions within hours.

Validation: Critically, the learned policies are validated against experimental data spanning joint kinematics, joint kinetics, ground reaction forces, and — most relevant to somatic AI — electromyography (EMG) recordings during walking and running. The model does not just look right; its muscle activations match real measured muscle activity.

Why it matters for somatic AI: This is the closest the mainstream motion AI field has come to SSIN's core premise. A model that generates movement through muscle activation, validated against real EMG, is operating at the layer where effort quality is produced — not at the downstream skeletal output. The open-source release (code and models on GitHub) makes muscle-level motion modelling accessible to movement researchers for the first time.

2. Personalised Muscle Control for Impaired Gait

Musculoskeletal Motion Imitation for Learning Personalised Exoskeleton Control Policy in Impaired Gait Choi, I., Park, I., Halilaj, E., & Kang, I. (Carnegie Mellon University). arXiv:2604.09431. https://arxiv.org/abs/2604.09431

The problem: Lower-limb exoskeletons that assist walking must be tuned to each individual — and for people with gait impairments, the tuning must account for their specific compensatory patterns. Conventional approaches require exhaustive per-user data collection or iterative on-body optimization, both slow and burdensome.

The approach: A device-agnostic framework combines physiologically plausible musculoskeletal simulation with reinforcement learning. Reference kinematics and joint moments are extracted from biomechanics data via OpenSim; the control policy observes future reference kinematics and outputs muscle activations imitating able-bodied locomotion. Crucially, the framework captures clinically observed compensatory strategies under targeted muscular deficits — providing a unified computational model of both healthy and pathological gait.

Why it matters: Beyond its rehabilitation application, this work demonstrates that muscle-level motion models can be personalised — adapted to an individual's specific neuromuscular condition, including atypical and compensatory patterns. For somatic AI, the personalisation of muscle-driven models is directly relevant: a practitioner's idiosyncratic movement vocabulary is, at the muscular level, their specific pattern of activation, compensation, and habit. A model that can represent individual muscular strategies is a model that can represent individual somatic identity.

3. Synthetic Data for Muscle-Level Learning: The Mouse Forelimb Result

Massively Parallel Imitation Learning of Mouse Forelimb Musculoskeletal Reaching Dynamics (Open-access, PMC12676374). https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12676374/

The contribution: An imitation learning framework simulates a dexterous forelimb reaching task with a musculoskeletal model in the MuJoCo physics environment, training at over 1 million steps per second through GPU acceleration. While the embodiment is a mouse forelimb rather than a human body, the result establishes that GPU-parallel muscle simulation can train motor policies at speeds previously impossible — the same computational unlock that MuscleMimic applies to the full human body.

Why it matters: Taken together with MuscleMimic and the CMU exoskeleton work, this paper confirms a clear field-wide development: the computational barrier that kept AI motion research at the skeletal level is falling. Muscle-level motion modelling — the level at which somatic effort quality is produced — is becoming computationally tractable at scale.

APA References

Choi, I., Park, I., Halilaj, E., & Kang, I. (2026). Musculoskeletal motion imitation for learning personalised exoskeleton control policy in impaired gait. arXiv:2604.09431. https://arxiv.org/abs/2604.09431

Li, C., Wang, C., Ziliotto, B., Simos, M., Kovecses, J., Durandau, G., & Mathis, A. (2026). Towards embodied AI with MuscleMimic: Unlocking full-body musculoskeletal motor learning at scale. arXiv:2603.25544. https://arxiv.org/abs/2603.25544